A STRUCTURED LEARNING APPROACH TO TEMPORAL RELATION EXTRACTION

Transcription

1 A STRUCTURED LEARNING APPROACH TO TEMPORAL RELATION EXTRACTION Qiang Ning, Zhili Feng, Dan Roth Computer Science University of Illinois, Urbana-Champaign & University of Pennsylvania 1

2 TOWARDS NATURAL LANGUAGE UNDERSTANDING Reasoning with respect to Time 2

3 UNDERSTANDING TIME IN TEXT Understanding time is key to understanding events Timeline construction (e.g., news stories, clinical records), time-slot filling, Q&A, causality analysis, pattern discovery, etc. Applications depend on two fundamental tasks Time expression extraction and normalization yesterday Time that is expressed explicitly Thursday after labor day time expressions in every 100 tokens (in TempEval3 datasets) Temporal relation extraction Time that is expressed implicitly A happens BEFORE/AFTER B 12 temporal relations in every 100 tokens (in TempEval3 datasets) 3

4 GRAPH REPRESENTATION OF TEMPORAL RELATIONS In Los Angeles that lesson was brought home today when tons of earth cascaded down a hillside, ripping two houses from their foundations. No one was hurt, but firefighters ordered the evacuation of nearby homes and said they'll monitor the shifting ground until March 23 rd. Five Relation types: ripping monitor Before; After; Include; Included; equal hurt cascaded ordered BEFORE INCLUDED 4

5 CHALLENGE I: STRUCTURE Structure of a temporal graph [Bramsen et al. 06; Chambers & Jurafsky 08l Do et. al. 12] Symmetry: A BEFORE B B AFTER A Transitivity: A BEFORE B + B BEFORE C A BEFORE C Relations are highly interrelated, but existing methods learn models by considering a single pair at a time. Existing methods Expectation ripping monitor ripping vs hurt ripping vs cascaded ripping vs monitor hurt cascaded ordered BEFORE INCLUDED 5

6 CHALLENGE II: MISSING RELATIONS Most of the relations are left unannotated Ground Truth Problems of existing approaches Addressing both challenges Structured Prediction Dealing with missing relations in the annotation. Provided Annotation (TempEval3) ripping monitor ripping monitor hurt hurt cascaded ordered cascaded ordered BEFORE INCLUDED MISSING Missing relations arise in three scenarios: The annotators did not look at a pair of events (e.g, long distance) The annotators could not decide among multiple options Annotators disagreements The annotation task is difficult if done at a single event pair level 6

7 EXISTING APPROACHES Local methods [1-4] Learn models or design rules that make pairwise decisions between each pair of events Global consistency (i.e., symmetry and transitivity) is not enforced Inconsistency may exist in local methods Local methods + Global Inference (L+I) [5-7] B A C Formulate the problem as an integer linear programming (ILP) over the entire graph, on top of pre-learnt local models Consistency guaranteed: structural requirements are added as declarative constraints to the ILP Performance improved: Local decisions may be corrected via global consideration L+I [1] Mani et al., ACL2006 [2] Chambers et al., ACL2007 [3] Bethard, ClearTK-TimeML: TempEval 2013 [4] Laokulrat et al., SEM2013 [5] Bramsen et al., EMNLP2006 [6] Chambers and Jurafsky, EMNLP2008 [7] Do et al., EMNLP2012 B A C Consistency is enforced via ILP 7

8 CHALLENGE I: CONSISTENT DECISION MAKING IS NOT SUFFICIENT Neither local methods nor L+I methods account for structural constraints in the learning phase. But information from other events is often necessary. tons of earth cascaded down a hillside, ripping two houses firefighters ordered the evacuation of nearby homes (What s the temporal relation between ripping and ordered? It s difficult to tell.) As a result, (ripping, ordered)=before cannot be supported given the local information, resulting in overfitting.. However, observing that (ripping, ordered)=before actually results from (ripping, cascaded)=included and (cascaded, ordered)=before, rather than the local text itself, supports better learning. ripping? ordered cascaded ordered ordered ripping ripping 8

9 PROPOSED APPROACH: INFERENCE-BASED TRAINING Local Training (Perceptron) For each x, y y = sgn(w T x) If y y Update w (x, y): feature and label for a single pair of events When learning from x, y, the algorithm is unaware of decisions with respect to other pairs. IBT (Structured Perceptron) For each (X, Y) Y = argmax Y C If Y Y W T X Update W X, Y : features and labels from a whole document Y C: Enforce consistency through constraint C. 9

10 PROPOSED APPROACH: INFERENCE-BASED TRAINING Inference step E Event node set, Y temporal label set I r (ij) Boolean variable for event pair (i,j) being relation r f r (ij) softmax score of event pair (i,j) being relation r r m temporal relations implied by r 1 and r 2 s.t. i, j, k E መI = arg min I ij E r Y f r ij I r (ij) I r ij = 1 r I r ij = I r ji I r1 ij + I r2 jk I rm ik 1 m Uniqueness Symmetry Generalized Transitivity 10

11 PROPOSED APPROACH: INFERENCE-BASED TRAINING Constraint-Driven Learning Make use of unannotated data Chang et al., Guiding semi-supervision with constraint-driven learning. ACL2007. Chang et al., Structured learning with constrained conditional models. Machine Learning

12 RESULTS (CHALLENGE I) When gold related pairs are known (TE3, Task C, Relation only) Enforcing constraints only at decision time. Enforcing constraints during learning System Method Precision Recall F1 UTTime [1] Local AP Local AP+ILP L+I SP+ILP S+I [1] Laokulrat et al., UTTime: Temporal relation classification using deep syntactic features, SEM

13 HOWEVER, REALISTICALLY When gold related pairs are NOT known (TE3, Task C) System Method Precision Recall F1 ClearTK [1] Local AP Local AP+ILP L+I SP+ILP S+I Performance drops significantly. Structured learning is not helping as much as previously in the presence of missing, vague relations Existing methods of handling vague relations are ineffective: Simply add vague to the temporal label set Train a classifier or design rules for vague vs. non-vague [1] Bethard, ClearTK-TimeML: A minimalist approach to TempEval

14 CHALLENGE II: MISSING RELATIONS Most of the relations are left unannotated Ground Truth Provided Annotation (TempEval3) ripping monitor ripping monitor hurt hurt cascaded ordered cascaded ordered BEFORE INCLUDED MISSING The annotation task is difficult if done at a single event pair level Some of the missing relations can be inferred Saturate the graph via symmetry and transitivity The vast majority, cannot 14

15 HANDLING VAGUE RELATIONS 1. Ignore vague labels during training Many vague pairs are not really vague but rather pairs that the annotators failed to look at. The imbalance between vague and non-vague relations makes it hard to learn a good vague classifier. The Vague relation is fundamentally different from other relation types. If (A, B) = BEFORE, then it s always BEFORE regardless of other events. But if (A, B) = VAGUE, the relation can change if more context is provided. 2. Apply post-filtering using KL divergence For each pair, we have a predicted distribution over possible relations. Compute the KL divergence of this distribution with the uniform distribution, and filter out predictions that have a low score. δ i = σ M m=1 f rm i log(mf rm i ), M=#labels, f r i =score for pair i. High similarity to the uniform distribution, δ i < t, implies unconfident prediction change decision to Vague

16 RESULTS (CHALLENGE II) When gold related pairs are NOT known (TE3, Task C) Apply the post-filtering method proposed above System Method Precision Recall F1 ClearTK [1] Local AP Local AP+ILP L+I SP+ILP S+I Applying post-filtering method for vague relations SP+ILP S+I CoDL+ILP S+I [1] Bethard, ClearTK-TimeML: A minimalist approach to TempEval

17 OVERALL RESULTS TempEval3 dataset is known to suffer from TLINK sparsity issues. Timebank-dense is another dataset with much denser TLINK annotations. Significant improvement over CAEVO, the previousely best system on Timebank-dense. System Method Precision Recall F1 ClearTK [1] Local CAEVO [2] L+I SP+ILP S+I CoDL+ILP S+I [1] Bethard, ClearTK-TimeML: A minimalist approach to TempEval 2013 [2] Chambers et al., Dense event ordering with a multi-pass architecture. TACL

18 CONCLUSION Thanks Identifying Temporal relations between events is a highly structured task This results also in low quality annotation (vague relations) This work shows that Using structured information during learning is important The structure can be exploited in an unsupervised way (via CoDL) to further improve results Vagueness is the result of lack of information rather than a concrete relation. KL-driven post-filtering is shown to be an effective way to treat vague relations. A lot more work is needed on temporal reasoning from text 18 18